Kriging Model combined with latin hypercube sampling for surrogate modeling of analog integrated circuit performance

The strong nonlinearity brought by the large circuit scale and more complicated physical/electrical models is making traditional response surface model (i.e. linear or quadratic polynomial) unsuitable for the surrogate-modeling of nowadays integrated circuits. Besides, the random-measurement-error-based analysis techniques and principles developed for traditional response surface model may be meaningless facing the deterministic data from circuit simulation experiments, which are not subject to random measurement errors essentially. Further, traditional response surface model can not mimic circuit behaviors in global process/designable parameter space. This paper proposes using Kriging model combined with Latin hypercube sampling to build surrogate model of circuit performance. We firstly introduce and compare the response surface modeling based on traditional design of experiments and the Kriging modeling combined with Latin hypercube sampling, and then apply both methods to circuit performance surrogate-modeling of an integrated operational amplifier. The result shows that Kriging model needs less sample points and provides 2times higher accuracy than quadratic response surface model does. Kriging modeling of circuit performance can be utilized to estimate parametric yield. Besides, it may facilitate the global optimization of parametric yield or circuit performance.